mltoys/pairwise_compare.py

#!/usr/bin/env python3
# pairwise_compare.py
import logging
import random
import torch
from torch import nn
from torch.nn import functional as F

# early pytorch device setup
DEVICE = torch.accelerator.current_accelerator() if torch.accelerator.is_available() else "cpu"

# Valves
DIMENSIONS = 1
TRAIN_STEPS = 25000
TRAIN_BATCHSZ = 16384
TRAIN_PROGRESS = 100
BATCH_LOWER = -512.0
BATCH_UPPER = 512.0
DO_VERBOSE_EARLY_TRAIN = False

def get_torch_info():
    log.info("PyTorch Version: %s", torch.__version__)
    log.info("HIP Version: %s", torch.version.hip)
    log.info("CUDA support: %s", torch.cuda.is_available())

    if torch.cuda.is_available():
        log.info("CUDA device detected: %s", torch.cuda.get_device_name(0))

    log.info("Using %s compute mode", DEVICE)

def set_seed(seed: int):
    random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)

# 1) Data: pairs (a, b) with label y = 1 if a > b else 0
def sample_batch(batch_size: int, low=BATCH_LOWER, high=BATCH_UPPER):
    a = (high - low) * torch.rand(batch_size, 1) + low
    b = (high - low) * torch.rand(batch_size, 1) + low

    # train for if a > b
    y = (a > b).float()

    # removed but left for my notes; it seems training for equality hurts classifing results that are ~eq
    # when trained only on "if a > b => y", the model produces more accurate results when classifing if things are equal (~.5 prob).
    # eq = (a == b).float()
    # y = gt + 0.5 * eq
    return a, b, y

# 2) Number "embedding" network: R -> R^d
class NumberEmbedder(nn.Module):
    def __init__(self, d=8):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(1, 16),
            nn.ReLU(),
            nn.Linear(16, d),
        )

    def forward(self, x):
        return self.net(x)

# 3) Comparator head: takes (ea, eb) -> logit for "a > b"
class PairwiseComparator(nn.Module):
    def __init__(self, d=8):
        super().__init__()
        self.embed = NumberEmbedder(d)
        self.head = nn.Sequential(
            nn.Linear(2 * d + 1, 16),
            nn.ReLU(),
            nn.Linear(16, 1),
        )

    def forward(self, a, b):
        ea = self.embed(a)
        eb = self.embed(b)
        delta_ab = a - b
        x = torch.cat([ea, eb, delta_ab], dim=-1)

        return self.head(x)  # logits
    
def training_entry():
    # all prng seeds to 0 for deterministic outputs durring testing
    # the seed should initialized normally otherwise
    set_seed(0)

    model = PairwiseComparator(d=DIMENSIONS).to(DEVICE)
    # opt = torch.optim.AdamW(model.parameters(), lr=2e-3)
    opt = torch.optim.Adadelta(model.parameters(), lr=1.0)

    # 4) Train
    for step in range(TRAIN_STEPS):
        a, b, y = sample_batch(TRAIN_BATCHSZ)
        a, b, y = a.to(DEVICE), b.to(DEVICE), y.to(DEVICE)

        logits = model(a, b)
        loss_fn = F.binary_cross_entropy_with_logits(logits, y)

        opt.zero_grad()
        loss_fn.backward()
        opt.step()

        if step <= TRAIN_PROGRESS and DO_VERBOSE_EARLY_TRAIN is True:
            with torch.no_grad():
                pred = (torch.sigmoid(logits) > 0.5).float()
                acc = (pred == y).float().mean().item()
            log.info(f"step={step:5d} loss={loss_fn.item():.7f} acc={acc:.7f}")
        elif step % TRAIN_PROGRESS == 0:
            with torch.no_grad():
                pred = (torch.sigmoid(logits) > 0.5).float()
                acc = (pred == y).float().mean().item()
            log.info(f"step={step:5d} loss={loss_fn.item():.7f} acc={acc:.7f}")

    # 5) Quick test: evaluate accuracy on fresh pairs
    with torch.no_grad():
        a, b, y = sample_batch(TRAIN_BATCHSZ)
        a, b, y = a.to(DEVICE), b.to(DEVICE), y.to(DEVICE)
        logits = model(a, b)
        pred = (torch.sigmoid(logits) > 0.5).float()
        errors = (pred != y).sum().item()
        acc = (pred == y).float().mean().item()
    log.info(f"Final test acc: {acc} errors: {errors}")

    # embed model depth into the model serialization
    torch.save({"state_dict": model.state_dict(), "d": DIMENSIONS}, "model.pth")
    log.info("Saved PyTorch Model State to model.pth")

def infer_entry():
    model_ckpt = torch.load("model.pth", map_location=DEVICE)
    model = PairwiseComparator(d=model_ckpt["d"]).to(DEVICE)
    model.load_state_dict(model_ckpt["state_dict"])
    model.eval()

    # sample pairs
    pairs = [(1, 2), (10, 3), (5, 5), (10, 35), (-64, 11), (300, 162), (2, 0), (2, 1), (3, 1), (4, 1), (3, 10),(30, 1), (0, 0), (-162, 237),
             (10, 20), (100, 30), (50, 50), (100, 350), (-640, 110), (30, -420), (200, 0), (92, 5), (30, 17), (42, 10), (30, 100),(30, 1), (0, 400), (-42, -42)]
    a = torch.tensor([[p[0]] for p in pairs], dtype=torch.float32, device=DEVICE)
    b = torch.tensor([[p[1]] for p in pairs], dtype=torch.float32, device=DEVICE)

    # sanity check before inference
    log.info(f"a.device: {a.device} model.device: {next(model.parameters()).device}")

    with torch.no_grad():
        probs = torch.sigmoid(model(a, b))

    for (x, y), p in zip(pairs, probs):
        log.info(f"P({x} > {y}) = {p.item():.3f}")

if __name__ == '__main__':
    import sys
    import os
    import datetime

    log = logging.getLogger(__name__)
    logging.basicConfig(filename='pairwise_compare.log', level=logging.INFO)
    log.info(f"Log opened {datetime.datetime.now()}")

    get_torch_info()

    name = os.path.basename(sys.argv[0])
    if name == 'train.py':
        training_entry()
    elif name == 'infer.py':
        infer_entry()
    else:
        # alt call patern
        # python3 pairwise_compare.py train
        # python3 pairwise_compare.py infer
        if len(sys.argv) > 1:
            mode = sys.argv[1].strip().lower()
            if mode == "train":
                training_entry()
            elif mode == "infer":
                infer_entry()
            else:
                log.error(f"Unknown operation: {mode}")
                log.error("Invalid call syntax, call script as \"train.py\" or \"infer.py\" or as pairwise_compare.py <mode> where mode is \"train\" or \"infer\"")
        else:
            log.error("Not enough arguments passed to script; call as train.py or infer.py or as pairwise_compare.py <mode> where mode is \"train\" or \"infer\"")

    log.info(f"Log closed {datetime.datetime.now()}")